Hosts and GPUs

Admin, Fleet, Hosts lists every machine running a node agent, with its GPUs.

The fleet host list, with per-GPU slot use and VRAM, RAM and disk bars.

Reading a host

Each host shows its status, when it last sent a heartbeat, its CPU, its storage, and a set of readiness checks.

The readiness checks are the useful part, because they report what the agent found, not what the configuration claims. The distinction matters. A host can be configured for NVENC and still have no working encoder.

Checks cover the render node, /dev/uinput write access, the managed home root, media reachability through the host firewall, and on NVIDIA the presence of the EGL and 32-bit GL components that a CUDA-only driver install leaves out.

A media_reachability warning names the scoped firewall rule the host is missing. See Ports and endpoints.

Readiness checks warn, they do not block

A failing check does not stop the host registering or accepting sessions. It tells you why something will probably not work. The exception is a host with no usable encoder, which will fail sessions rather than pretend.

GPU capacity and admission

Each GPU shows its encode slots and its VRAM, used and free.

A session is admitted to a GPU when two conditions hold.

Encode slots. This is the real reservation, and it is race-safe. A GPU has a finite number of concurrent encodes it can sustain.

Free VRAM. A veto, not a reservation. If the GPU’s most recent memory sample shows less free than QUASAR_VRAM_MIN_FREE_MB (1024 MB by default), the GPU is skipped. Very recent launches the sample has not caught up with are debited from the figure.

The veto abstains rather than blocking when the telemetry is missing, stale or structurally implausible. An integrated GPU whose entire memory carve-out sits below the floor would otherwise reject everything. Unknown telemetry is never allowed to reduce availability.

Set QUASAR_VRAM_MIN_FREE_MB=0 to turn the veto off and admit on slots alone.

Declared per-app VRAM is gone

Apps used to carry a declared VRAM figure. It was never enforceable, because there is no VRAM cgroup controller to enforce it with, and in practice it was an admin-typed guess nothing reconciled. Live telemetry replaced it.

The two capacity errors

no_host_available means nothing in the fleet could ever serve the request. Usually every host is offline. Retrying does not help.

capacity_exhausted means hosts are up but currently full. Retrying does help, once something ends.

Per-host settings

Each host has a Settings page that overrides its environment configuration without editing .env and restarting anything by hand. Precedence runs: per-host override, then agent environment, then catalog default.

Settings come in two classes.

Live settings apply to the next session on that host. ABR mode, container knobs and most tuning are live.

Restart settings need the agent to restart, which drops any live sessions on that host. The encoder, the render node and the CUDA device are in this class. Quasar tells you before it does it.

Render node selection

On a machine with more than one GPU, QUASAR_RENDER_NODE picks which one the compositor and encoder use.

Prefer a /dev/dri/by-path/ path over /dev/dri/renderD128. The numbered names are assigned in probe order and can move between reboots. The by-path names are tied to the PCI address and do not.

For NVENC on a multi-GPU box, QUASAR_CUDA_DEVICE selects the CUDA device index as well.

Per-session GPU routing does not exist yet

Placement chooses a host, and the host’s configuration chooses a GPU. You cannot currently route a specific session to a specific GPU on a multi-GPU host.

Placement across hosts

QUASAR_PLACEMENT_POLICY on the control plane:

Value	Behaviour
spread	Default. Distributes sessions across hosts.
least-loaded	Prefers the host with the most free capacity.
locality	Prefers the host that already holds the user’s home directory.

Use locality if home directories are host-local rather than on shared storage. Otherwise a user’s saves are on whichever host happened to run the session.

Draining and removing

Drain stops a host accepting new sessions while letting existing ones finish. Do this before maintenance.

Once it is empty, stop the agent on that machine and forget the host here.

The console

Each host has a console page for local-display mode, where the host renders to a physically attached screen rather than to a browser. It needs additional privileges and a separate compose overlay, and it is not part of the default deployment.

Storage

Admin, Fleet, Storage shows the managed home directories across your fleet.

Each host can have its own managed-home root. It must be genuinely shared storage if you want saves to follow users between machines; otherwise use locality placement to keep a user near the host holding their files. The readiness check reports whether the root is visible and writable from the agent’s namespace.

Changing a storage root does not migrate data

Pointing a host at a different directory makes new launches use that directory. Existing homes remain at the old path until you move them deliberately. To a user, changing the root without migrating data looks like their saves vanished.